Artificial intelligence in gynecologic cancers: Current status and future challenges - A systematic review

The rapid expansion of computational diagnostic tools has transformed modern medical practice. No prior work had resolved the specific landscape of these technologies within the field of gynecologic oncology. It was already known that automated systems offer potential benefits for patient care. That uncertainty drove a need to categorize existing literature systematically. Prior research has shown that diagnostic accuracy varies across different cancer types. This gap motivated a comprehensive assessment of current progress. Researchers have observed a shift toward using complex algorithms for clinical decision support. The field currently lacks a unified framework for evaluating these diverse technological approaches.

The aim of this study was to elucidate the current state of research regarding computational diagnostic tools in gynecologic cancers. This investigation sought to map the landscape of existing literature published over the last decade. Researchers intended to identify which specific cancer types receive the most attention in the current technological climate. The study also aimed to categorize the types of data inputs commonly utilized in these predictive models. By examining these factors, the authors hoped to highlight the strengths and weaknesses of current methodologies. This work addresses the need for a clearer understanding of how these tools perform in clinical settings. The authors also sought to identify common challenges that limit the widespread adoption of these systems. This review provides a foundation for evaluating future advancements in the field.

The review approach involved searching three major databases for relevant research published between 2010 and 2020. Investigators excluded genomic studies, molecular research, and specific automated screening procedures like digital colposcopy. This process yielded 71 eligible papers from an initial pool of 1632 records. The team categorized inputs into imaging-based and value-based data streams. Imaging sources included magnetic resonance imaging, computed tomography, ultrasound, and hysteroscopy. Value-based inputs comprised patient demographics, blood test results, tumor markers, and pathological indices. Researchers assessed model quality using standard statistical performance indicators. The team opted against quantitative synthesis due to the high variability observed across the collected publications.

Key findings from the literature indicate that 34 studies focused on cervical cancer, while 21 addressed ovarian cancer. Endometrial cancer and uterine sarcoma were represented by 13 and three studies, respectively. Imaging data served as the input for 35 studies, whereas 36 studies utilized value-based information. The median dataset size across all analyzed works was 214 cases. A total of 90% of the studies, or 64 papers, relied on datasets containing fewer than 1000 cases. Cervical cancer research frequently prioritized prognostic outcomes over diagnostic classification. Conversely, ovarian cancer investigations primarily emphasized diagnostic accuracy. The authors observed significant heterogeneity, which precluded a formal meta-analysis of the reported performance metrics.

The authors suggest that cervical cancer remains the most frequently studied malignancy in this domain. Prognostic modeling appears more prevalent for cervical cases compared to ovarian cancer studies. Diagnostic classification serves as the primary focus for ovarian cancer investigations. Limited evidence exists regarding the effectiveness of these models for endometrial cancer and uterine sarcoma. Small sample sizes frequently hinder the robustness of current predictive systems. The absence of external validation datasets represents a major barrier to clinical implementation. Heterogeneity across studies prevents a formal quantitative synthesis of existing performance metrics. Future efforts should prioritize larger, multi-center datasets to improve model reliability and generalizability.